ConciseMode

practice · published · latency · local-llm · agents

Output generation latency, not reasoning, is the bottleneck on slow inference paths (e.g. local LLMs at single-digit tokens/second). Conversational pleasantries, hedging, and verbose recaps cost real wall-clock time without adding signal. ConciseMode defines three selectable tiers of terseness an agent can be asked to operate under.

Core invariant across all levels: code, diffs, tool parameters, and command syntax are always preserved 100% verbatim. ConciseMode trims narration, never content.

---

Level 1 — Concise Lite (Professional Terse)

• Drop opening pleasantries ("Certainly!", "I'll be happy to help...") and speculative closing remarks.
• Keep full grammar and complete-sentence explanations.
• Default tier for interactive user pairing — the reader still wants prose, just without filler.

Level 2 — Concise Standard (Telegraphic / Core Caveman)

• Strip articles, connective phrases, and grammatical filler.
• Structure findings as dense fragments: Action -> Finding -> Patch.
• Use for autonomous subagents and fast canary/benchmark sweeps where no human reads the stream live.

Level 3 — Concise Ultra (Extreme Shorthand / Zero-Fluff)

• Output confined to essential diffs, command invocations, and single-line status confirmations, e.g. PASS: 14 tests, built bin/app.
• Zero narrative text.
• Use only on resource-constrained or low-TPS local hardware where every output token has measurable latency cost.

---

Enforcing a Default Tier

Naming the tiers is not enough — an agent only runs under one when something pins it there. The doc reference alone (@docs/practices/ConciseMode.md) is descriptive, not a directive.

• CLAUDE.md directive (recommended default): add an explicit line naming the tier, e.g. Operate at Concise Lite (@docs/practices/ConciseMode.md) unless told otherwise. This loads into every session's system context automatically and is scoped per project, but stays advisory — a long session can still drift from it.
• Output style (settings.json outputStyle): bakes the tier into the harness-level system prompt instead of a doc reference. Stronger and harder to drift from, but global to the session/install rather than toggleable per task.

Pick the CLAUDE.md directive first; reach for an output style only if drift is observed in practice.

Operational Pairing Format (Deployment / Live-Host Sessions)

Live deployment and infrastructure-provisioning pairing sessions (SSH sessions, remote debugging, "is it actually deployed" checks) have a reporting need distinct from the three narrative tiers above: not just less prose, but a fixed structure so an operator can scan the response for the one fact that matters — usually a gap. This emerged from a live infrastructure pairing retrospective: switching to a strict bullet structure measurably improved operator alignment and cycle speed over both normal prose and the generic terse tiers above.

When operating in this mode, structure each response as 3–7 one-line bullets, in this order, omitting any category with nothing to report:

• Fact / Command — the exact command or check run.
• Status — exit code or direct observation (not an interpretation).
• Gaps / Discrepancies — anything that doesn't match what local state predicted (this is the category most narrative styles bury under reassurance — see the optional deployment-transparency practice when the project has remote deployment).
• Next action — the single recommended next step.

This format composes with the tiers above rather than replacing them: use it specifically while probing or reporting on remote/deployed state; drop back to Level 1–3 narrative for everything else in the same session. Invoke via /mode (see commands/mode.md) or by naming it directly in a pairing-session directive, e.g. "use operational pairing format for the rest of this deploy."

Pairing with Output Distillation

ConciseMode governs what the agent writes; harnez distill (see issue 066) governs what the agent reads back from tool output. The two compound: distillation keeps noisy command output out of context, ConciseMode keeps the agent's own narration short. Neither substitutes for the other — apply both on constrained hardware.